A generic scientific machine learning framework for fruit and vegetable quality prediction

Accurately predicting the postharvest quality of fruits and vegetables (F&V) is a critical challenge for supply chain management and consumer satisfaction. Traditional knowledge-driven kinetic models have good interpretability but often struggle to account for diverse pre-storage and environmental factors, particularly for users without experience in kinetic modeling. In contrast, purely data-driven approaches, such as neural networks, require large datasets and are unable to leverage domain knowledge to improve data efficiency. In this study, we propose a conceptual framework for Scientific Machine Learning (SciML) that integrates domain knowledge with data-driven neural networks, providing a hybrid approach for F&V quality prediction. A decision tree is included to guide model selection based on dataset characteristics, the availability of prior knowledge, and prediction goals. As a proof of concept, we first conducted a study using synthetically generated quality decay data, demonstrating that incorporating kinetic knowledge directly into the neural network improves predictive accuracy and temporal extrapolation. Subsequently, a real-world case study predicting the overall visual quality (OVQ) of greenhouse tomatoes further confirmed the advantages of hybrid knowledge-data integration. Overall, this research demonstrates that SciML offers a third modeling alternative that balances data-driven flexibility and knowledge-driven interpretability. While not intended to replace traditional approaches, SciML can enrich the toolbox available to researchers and practitioners for F&V quality modeling and, more broadly, for food quality prediction. Human judgment remains essential in selecting and applying the appropriate SciML model to ensure an effective balance among prediction accuracy on limited datasets, knowledge incorporation for improved extrapolation, computational efficiency, and interpretability. • A conceptual framework is proposed for Scientific Machine Learning (SciML). • The framework supports postharvest quality prediction of fruits and vegetables. • Standardized definitions and implementation guidelines are established. • Synthetic and real-world case studies validate the framework’s effectiveness.